Plain-English definitions for the terms used in the ROC framework. Built for leadership teams that need to make AI investment decisions together.
An AI system that uses repeated inference plus tools, memory, and decision logic to complete multi-step tasks. Agents consume more compute per useful output than single-shot inference, but can replace much larger blocks of human work.
The portion of a future outcome that is attributable to the AI compute investment, as opposed to other investments (sales, GTM, design, distribution). Prevents AI from being credited with revenue that would have happened anyway.
The portion of saved employee time that converts into productive economic value. Most organizations land between 25% and 60%. A 100% capture rate assumes every saved hour becomes real economic output, which almost never happens.
The processing capacity used to run AI workloads. Sourced from GPUs, TPUs, custom AI chips, cloud APIs, hosted models, or internal clusters. Compute powers both model creation (training, fine-tuning) and model usage (inference, agentic workflows).
Revenue minus variable cost, expressed as a percentage. Used to convert incremental revenue into incremental economic value, so high-volume / low-margin revenue is not over-credited.
Total compute (and supporting) cost divided by useful outputs produced. The single most important efficiency metric in ROC, far more durable than cost per token.
The value the organization would have captured without the AI investment. Subtracted from total expected value so only the incremental value is counted.
Raw compute multiplied by five operating factors: utilization, fungibility, reliability, price-performance, and model/workload efficiency. Two organizations with the same raw spend can produce wildly different effective compute.
Adapting an existing base model to a narrower task or domain using additional training data. Less expensive than training from scratch, but compute cost scales with parameter count and dataset size.
The ability to move workloads across vendors, models, chips, or clouds. High fungibility lowers cost-per-output by enabling routing to the best available capacity at any moment.
Graphics Processing Unit. Originally for graphics, GPUs have become the dominant accelerator for AI workloads because their parallel architecture is well-suited to the matrix operations at the heart of model training and inference.
Running a trained model to produce outputs: answering a question, drafting an email, summarizing a document. Inference scales with usage and is typically the largest compute line for production AI.
The return on the next incremental dollar of compute in a specific bucket. Marginal ROC is the decision-making metric. It tells you where the next dollar should go, regardless of how total ROC looks today.
Future cash flows discounted to their value today. ROC uses NPV on both sides of the ratio so dollars committed now and dollars created over a multi-year horizon are compared on a consistent basis.
The value of having enough flexible capacity to serve upside demand without stranding cost on the downside. ROC computes option value as a scenario-weighted expectation across user-supplied demand scenarios.
The time at which cumulative value created by AI compute equals cumulative compute cost. ROC reports payback alongside the multiple to flag programs that look good in aggregate but won't recoup their investment inside the planning horizon.
Output per dollar across a provider mix. Distinct from headline price: a more expensive model with higher quality can deliver better price-performance for the right workload.
Risk adjustment applied to capability initiatives whose outcomes are uncertain. Lets the model count expected value without pretending uncertain futures are guaranteed.
A unit of input or output for a language model, roughly a word fragment. Pricing and capacity are often measured in tokens, but tokens are an input metric. The corresponding outcome metric is the useful output.
Tensor Processing Unit. Google's custom AI accelerator, optimized for the tensor math that dominates deep learning. TPUs offer different price-performance and availability tradeoffs than general-purpose GPUs.
The process of creating or improving a model from data. Frontier-scale training runs are among the largest compute commitments in the industry and are usually amortized across the lifetime of the resulting model.
A business event that the AI system produces or enables: a resolved support ticket, an accepted code change, a qualified lead, a reviewed contract. ROC prefers cost per useful output over cost per token as the operating metric.
The portion of purchased compute that is productively used. Low utilization is one of the most common sources of underperforming AI economics.